clanka 0.8.6 → 0.10.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (96) hide show
  1. package/dist/Acp.d.ts +6 -1
  2. package/dist/Acp.d.ts.map +1 -1
  3. package/dist/Acp.js +34 -3
  4. package/dist/Acp.js.map +1 -1
  5. package/dist/Acp.test.js +178 -2
  6. package/dist/Acp.test.js.map +1 -1
  7. package/dist/AcpImage.test.js +7 -1
  8. package/dist/AcpImage.test.js.map +1 -1
  9. package/dist/Agent.d.ts +82 -8
  10. package/dist/Agent.d.ts.map +1 -1
  11. package/dist/Agent.js +556 -113
  12. package/dist/Agent.js.map +1 -1
  13. package/dist/Agent.test.js +86 -5
  14. package/dist/Agent.test.js.map +1 -1
  15. package/dist/AgentCallerConfig.test.d.ts +2 -0
  16. package/dist/AgentCallerConfig.test.d.ts.map +1 -0
  17. package/dist/AgentCallerConfig.test.js +335 -0
  18. package/dist/AgentCallerConfig.test.js.map +1 -0
  19. package/dist/AgentEnvironment.d.ts +76 -0
  20. package/dist/AgentEnvironment.d.ts.map +1 -0
  21. package/dist/AgentEnvironment.js +119 -0
  22. package/dist/AgentEnvironment.js.map +1 -0
  23. package/dist/AgentEnvironment.test.d.ts +2 -0
  24. package/dist/AgentEnvironment.test.d.ts.map +1 -0
  25. package/dist/AgentEnvironment.test.js +304 -0
  26. package/dist/AgentEnvironment.test.js.map +1 -0
  27. package/dist/AgentExecutor.d.ts +342 -183
  28. package/dist/AgentExecutor.d.ts.map +1 -1
  29. package/dist/AgentExecutor.js +607 -127
  30. package/dist/AgentExecutor.js.map +1 -1
  31. package/dist/AgentExecutor.test.js +287 -31
  32. package/dist/AgentExecutor.test.js.map +1 -1
  33. package/dist/AgentExecutorImage.test.js +5 -1
  34. package/dist/AgentExecutorImage.test.js.map +1 -1
  35. package/dist/AgentImage.test.js +15 -2
  36. package/dist/AgentImage.test.js.map +1 -1
  37. package/dist/AgentOutput.d.ts +70 -5
  38. package/dist/AgentOutput.d.ts.map +1 -1
  39. package/dist/AgentOutput.js +65 -0
  40. package/dist/AgentOutput.js.map +1 -1
  41. package/dist/AgentSkills.d.ts +37 -0
  42. package/dist/AgentSkills.d.ts.map +1 -1
  43. package/dist/AgentSkills.js +96 -32
  44. package/dist/AgentSkills.js.map +1 -1
  45. package/dist/AgentSkills.test.js +14 -2
  46. package/dist/AgentSkills.test.js.map +1 -1
  47. package/dist/AgentSubagents.test.d.ts +2 -0
  48. package/dist/AgentSubagents.test.d.ts.map +1 -0
  49. package/dist/AgentSubagents.test.js +494 -0
  50. package/dist/AgentSubagents.test.js.map +1 -0
  51. package/dist/AgentTools.d.ts +356 -14
  52. package/dist/AgentTools.d.ts.map +1 -1
  53. package/dist/AgentTools.js +214 -35
  54. package/dist/AgentTools.js.map +1 -1
  55. package/dist/OutputFormatter.d.ts.map +1 -1
  56. package/dist/OutputFormatter.js +19 -2
  57. package/dist/OutputFormatter.js.map +1 -1
  58. package/dist/ToolkitRenderer.d.ts.map +1 -1
  59. package/dist/ToolkitRenderer.js +6 -1
  60. package/dist/ToolkitRenderer.js.map +1 -1
  61. package/dist/bin/{Effect-rADALOOD.mjs → Effect-2jPlJ8eu.mjs} +43 -38
  62. package/dist/bin/Effect-2jPlJ8eu.mjs.map +1 -0
  63. package/dist/bin/{McpClient-NazR2OwL.mjs → McpClient-1W508adc.mjs} +3 -3
  64. package/dist/bin/{McpClient-NazR2OwL.mjs.map → McpClient-1W508adc.mjs.map} +1 -1
  65. package/dist/bin/{OutputFormatter-BVHbgF2b.mjs → OutputFormatter-BFAv48k9.mjs} +17 -5
  66. package/dist/bin/{OutputFormatter-BVHbgF2b.mjs.map → OutputFormatter-BFAv48k9.mjs.map} +1 -1
  67. package/dist/bin/{Schema-C7NXAOWA.mjs → Schema-D-7VuCiC.mjs} +2 -2
  68. package/dist/bin/{Schema-C7NXAOWA.mjs.map → Schema-D-7VuCiC.mjs.map} +1 -1
  69. package/dist/bin/{SemanticSearch-4Jes3jCl.mjs → SemanticSearch-Cb8aqvGv.mjs} +4 -4
  70. package/dist/bin/{SemanticSearch-4Jes3jCl.mjs.map → SemanticSearch-Cb8aqvGv.mjs.map} +1 -1
  71. package/dist/bin/{Stream-CSLA1_GD.mjs → Stream-DEu2tfCV.mjs} +12 -3
  72. package/dist/bin/Stream-DEu2tfCV.mjs.map +1 -0
  73. package/dist/bin/cli.mjs +13333 -12470
  74. package/dist/bin/cli.mjs.map +1 -1
  75. package/package.json +1 -1
  76. package/src/Acp.test.ts +279 -3
  77. package/src/Acp.ts +34 -4
  78. package/src/AcpImage.test.ts +7 -1
  79. package/src/Agent.test.ts +109 -4
  80. package/src/Agent.ts +874 -140
  81. package/src/AgentCallerConfig.test.ts +522 -0
  82. package/src/AgentEnvironment.test.ts +504 -0
  83. package/src/AgentEnvironment.ts +161 -0
  84. package/src/AgentExecutor.test.ts +414 -37
  85. package/src/AgentExecutor.ts +1028 -191
  86. package/src/AgentExecutorImage.test.ts +5 -1
  87. package/src/AgentImage.test.ts +21 -3
  88. package/src/AgentOutput.ts +91 -1
  89. package/src/AgentSkills.test.ts +14 -2
  90. package/src/AgentSkills.ts +117 -34
  91. package/src/AgentSubagents.test.ts +824 -0
  92. package/src/AgentTools.ts +377 -53
  93. package/src/OutputFormatter.ts +19 -2
  94. package/src/ToolkitRenderer.ts +7 -1
  95. package/dist/bin/Effect-rADALOOD.mjs.map +0 -1
  96. package/dist/bin/Stream-CSLA1_GD.mjs.map +0 -1
package/package.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "name": "clanka",
3
3
  "type": "module",
4
- "version": "0.8.6",
4
+ "version": "0.10.0",
5
5
  "publishConfig": {
6
6
  "access": "public"
7
7
  },
package/src/Acp.test.ts CHANGED
@@ -1,9 +1,13 @@
1
1
  import { assert, describe, it } from "@effect/vitest"
2
+ import * as NodeServices from "@effect/platform-node/NodeServices"
3
+ import * as Context from "effect/Context"
2
4
  import * as Effect from "effect/Effect"
5
+ import * as FileSystem from "effect/FileSystem"
3
6
  import * as Fiber from "effect/Fiber"
4
7
  import * as Layer from "effect/Layer"
5
8
  import * as Option from "effect/Option"
6
9
  import * as Stream from "effect/Stream"
10
+ import * as FetchHttpClient from "effect/http/FetchHttpClient"
7
11
  import * as Prompt from "effect/ai/Prompt"
8
12
  import * as KeyValueStore from "effect/persistence/KeyValueStore"
9
13
  import * as LanguageModel from "effect/ai/LanguageModel"
@@ -24,8 +28,14 @@ type Message = {
24
28
 
25
29
  const capabilities = new AgentExecutor.Capabilities({
26
30
  toolsDts: "",
27
- agentsMd: Option.none(),
31
+ environment: new AgentExecutor.ExecutionEnvironment({
32
+ cwd: "/",
33
+ baseline: [],
34
+ seen: [],
35
+ }),
36
+ instructions: Option.none(),
28
37
  supportsSearch: false,
38
+ tools: [],
29
39
  skills: [],
30
40
  })
31
41
 
@@ -35,6 +45,16 @@ const executor = AgentExecutor.AgentExecutor.of({
35
45
  executeUnsafe: () => Effect.die("executeUnsafe not implemented"),
36
46
  })
37
47
 
48
+ // A fresh executor per server, like the CLI builds one per session.
49
+ const localExecutor = (directory: string) =>
50
+ Layer.build(
51
+ AgentExecutor.layerLocal({ directory }).pipe(
52
+ Layer.provide(Layer.mergeAll(NodeServices.layer, FetchHttpClient.layer)),
53
+ ),
54
+ ).pipe(
55
+ Effect.map((context) => Context.get(context, AgentExecutor.AgentExecutor)),
56
+ )
57
+
38
58
  const makeServer = (
39
59
  streamText: Parameters<typeof LanguageModel.make>[0]["streamText"],
40
60
  agentExecutor = executor,
@@ -99,6 +119,25 @@ const assistantSays = (text: string) =>
99
119
  { type: "text-end", id: "1" },
100
120
  )
101
121
 
122
+ // A model that runs `script` once, then answers "done".
123
+ const runsScript = (script: string) => {
124
+ let calls = 0
125
+ return () =>
126
+ calls++ === 0
127
+ ? parts({
128
+ type: "tool-call",
129
+ id: "call-1",
130
+ name: "execute",
131
+ params: { script },
132
+ })
133
+ : assistantSays("done")
134
+ }
135
+
136
+ const message = (role: "user" | "assistant", text: string) =>
137
+ Prompt.makeMessage(role, {
138
+ content: [Prompt.makePart("text", { text })],
139
+ })
140
+
102
141
  const withStore = <A, E, R>(effect: Effect.Effect<A, E, R>) =>
103
142
  effect.pipe(
104
143
  Effect.provide(
@@ -194,6 +233,8 @@ describe("Acp", () => {
194
233
  .find((update) => update.sessionUpdate === "usage_update"),
195
234
  {
196
235
  sessionUpdate: "usage_update",
236
+ used: 120,
237
+ size: 236_000,
197
238
  usage: {
198
239
  inputTokens: 84,
199
240
  outputTokens: 12,
@@ -244,6 +285,8 @@ describe("Acp", () => {
244
285
  .find((update) => update.sessionUpdate === "usage_update"),
245
286
  {
246
287
  sessionUpdate: "usage_update",
288
+ used: 120,
289
+ size: 236_000,
247
290
  usage: {
248
291
  inputTokens: 120,
249
292
  outputTokens: 12,
@@ -425,8 +468,20 @@ describe("Acp", () => {
425
468
  : assistantSays(summary),
426
469
  {
427
470
  ...executor,
428
- execute: ({ onSubagent }) =>
429
- Stream.fromEffect(onSubagent(prompt)),
471
+ execute: ({ subagents, environment }) =>
472
+ Stream.fromEffect(
473
+ subagents.delegate({
474
+ prompt,
475
+ background: false,
476
+ environment: new AgentExecutor.EnvironmentUpdate({
477
+ environment,
478
+ update: Option.none(),
479
+ }),
480
+ }),
481
+ ).pipe(
482
+ Stream.map((result) => result.output ?? ""),
483
+ Stream.orDie,
484
+ ),
430
485
  },
431
486
  )
432
487
  const created = yield* server.request(1, "session/new", {
@@ -468,6 +523,126 @@ describe("Acp", () => {
468
523
  ),
469
524
  )
470
525
 
526
+ it.effect(
527
+ "resuming a finished subagent does not create its tool call twice",
528
+ () =>
529
+ withStore(
530
+ Effect.scoped(
531
+ Effect.gen(function* () {
532
+ let childId: string | undefined
533
+ let executions = 0
534
+ const server = yield* makeServer(
535
+ ({ prompt }) => {
536
+ if (JSON.stringify(prompt.content).includes("TASK-CHILD")) {
537
+ const resumed = JSON.stringify(prompt.content).includes(
538
+ "FOLLOW-UP",
539
+ )
540
+ return assistantSays(
541
+ resumed ? "SECOND-RESULT" : "FIRST-RESULT",
542
+ )
543
+ }
544
+ const seen = JSON.stringify(prompt.content)
545
+ if (seen.includes("SECOND-RESULT")) return assistantSays("done")
546
+ if (
547
+ executions === 0 ||
548
+ (executions === 1 && seen.includes("FIRST-RESULT"))
549
+ ) {
550
+ return parts({
551
+ type: "tool-call",
552
+ id: `call-${executions}`,
553
+ name: "execute",
554
+ params: { script: "await work()" },
555
+ })
556
+ }
557
+ return assistantSays("waiting")
558
+ },
559
+ {
560
+ ...executor,
561
+ // The first script starts a background child; the second
562
+ // resumes it once its first result has been delivered.
563
+ execute: ({ subagents, environment }) =>
564
+ Stream.fromEffect(
565
+ executions++ === 0
566
+ ? subagents
567
+ .delegate({
568
+ prompt: "TASK-CHILD",
569
+ background: true,
570
+ environment: new AgentExecutor.EnvironmentUpdate({
571
+ environment,
572
+ update: Option.none(),
573
+ }),
574
+ })
575
+ .pipe(
576
+ Effect.map((result) => {
577
+ childId = result.id
578
+ return "started"
579
+ }),
580
+ )
581
+ : subagents
582
+ .messageAgent({ id: childId!, message: "FOLLOW-UP" })
583
+ .pipe(Effect.as("resumed")),
584
+ ).pipe(Stream.orDie),
585
+ },
586
+ )
587
+ const created = yield* server.request(1, "session/new", {
588
+ cwd: "/tmp",
589
+ })
590
+ const sessionId: string = created.result.sessionId
591
+ yield* server.request(2, "session/prompt", {
592
+ sessionId,
593
+ prompt: [{ type: "text", text: "delegate and resume" }],
594
+ })
595
+
596
+ assert.strictEqual(executions, 2, "the child was resumed")
597
+ const tools = server
598
+ .updates(sessionId)
599
+ .filter(
600
+ (u) =>
601
+ u.sessionUpdate.startsWith("tool_call") &&
602
+ u.toolCallId.startsWith("subagent-"),
603
+ )
604
+ const creations = tools
605
+ .filter((u) => u.sessionUpdate === "tool_call")
606
+ .map((u) => u.toolCallId)
607
+ assert.deepStrictEqual(
608
+ creations,
609
+ [...new Set(creations)],
610
+ "ACP tool call IDs must be unique within the session",
611
+ )
612
+ for (const u of tools) {
613
+ assert.include(
614
+ creations,
615
+ u.toolCallId,
616
+ "updates refer to a created tool call",
617
+ )
618
+ }
619
+
620
+ const firstDone = tools.findIndex(
621
+ (u) => u.status === "completed" && u.rawOutput === "FIRST-RESULT",
622
+ )
623
+ const secondDone = tools.findIndex(
624
+ (u) =>
625
+ u.status === "completed" && u.rawOutput === "SECOND-RESULT",
626
+ )
627
+ assert.isAbove(firstDone, -1, "first result shown")
628
+ assert.isAbove(secondDone, firstDone, "second result shown")
629
+ assert.isTrue(
630
+ tools
631
+ .slice(firstDone + 1, secondDone)
632
+ .some((u) => u.status === "in_progress"),
633
+ "the resumed turn is shown as running",
634
+ )
635
+ assert.deepStrictEqual(tools[secondDone].content, [
636
+ {
637
+ type: "content",
638
+ content: { type: "text", text: "SECOND-RESULT" },
639
+ },
640
+ ])
641
+ }),
642
+ ),
643
+ ),
644
+ )
645
+
471
646
  it.effect("reports canonical state for new and loaded sessions", () =>
472
647
  withStore(
473
648
  Effect.gen(function* () {
@@ -621,6 +796,107 @@ describe("Acp", () => {
621
796
  ),
622
797
  )
623
798
 
799
+ it.live("keeps the working directory across a session reload", () =>
800
+ withStore(
801
+ Effect.gen(function* () {
802
+ const fs = yield* FileSystem.FileSystem
803
+ const cwd = yield* fs.makeTempDirectoryScoped()
804
+ const sessionId = yield* Effect.scoped(
805
+ Effect.gen(function* () {
806
+ const server = yield* makeServer(
807
+ runsScript(`await bash({ command: "mkdir sub" })
808
+ await changeDirectory("sub")`),
809
+ yield* localExecutor(cwd),
810
+ )
811
+ const created = yield* server.request(1, "session/new", { cwd })
812
+ const sessionId: string = created.result.sessionId
813
+ yield* server.request(2, "session/prompt", {
814
+ sessionId,
815
+ prompt: [{ type: "text", text: "change directory" }],
816
+ })
817
+ return sessionId
818
+ }),
819
+ )
820
+
821
+ const output = yield* Effect.scoped(
822
+ Effect.gen(function* () {
823
+ const server = yield* makeServer(
824
+ runsScript(
825
+ `console.log("PWD:" + (await bash({ command: "pwd" })).trim())`,
826
+ ),
827
+ yield* localExecutor(cwd),
828
+ )
829
+ yield* server.request(1, "session/load", { sessionId, cwd })
830
+ yield* server.request(2, "session/prompt", {
831
+ sessionId,
832
+ prompt: [{ type: "text", text: "print it" }],
833
+ })
834
+ return server
835
+ .updates(sessionId)
836
+ .filter((u) => u.sessionUpdate === "tool_call_update")
837
+ .map((u) => String(u.rawOutput))
838
+ .join("\n")
839
+ }),
840
+ )
841
+ assert.match(output, /PWD:\S*\/sub$/m)
842
+ }).pipe(Effect.scoped, Effect.provide(NodeServices.layer)),
843
+ ),
844
+ )
845
+
846
+ it.effect(
847
+ "hides environment updates and sub-agent results during replay",
848
+ () =>
849
+ withStore(
850
+ Effect.scoped(
851
+ Effect.gen(function* () {
852
+ const kvs = yield* KeyValueStore.KeyValueStore
853
+ const store = KeyValueStore.toSchemaStore(
854
+ KeyValueStore.prefix(kvs, "session-"),
855
+ Acp.SessionRecord,
856
+ )
857
+ const sessionId = "injected-session"
858
+ yield* store.set(
859
+ sessionId,
860
+ new Acp.SessionRecord({
861
+ cwd: "/tmp",
862
+ model: "test:model",
863
+ thoughtLevel: "medium",
864
+ history: Prompt.fromMessages([
865
+ message("user", "Typed by the user"),
866
+ message(
867
+ "user",
868
+ "<environment-update>\nPKG-RULES\n</environment-update>",
869
+ ),
870
+ message(
871
+ "user",
872
+ '<subagent id="1" status="completed">\nCHILD-RESULT\n</subagent>',
873
+ ),
874
+ message("assistant", "Reply"),
875
+ ]),
876
+ }),
877
+ )
878
+
879
+ const server = yield* makeServer(() => assistantSays("unused"))
880
+ const loaded = yield* server.request(1, "session/load", {
881
+ sessionId,
882
+ cwd: "/tmp",
883
+ })
884
+ assert.isUndefined(loaded.error)
885
+ assert.deepStrictEqual(server.updates(sessionId), [
886
+ {
887
+ sessionUpdate: "user_message_chunk",
888
+ content: { type: "text", text: "Typed by the user" },
889
+ },
890
+ {
891
+ sessionUpdate: "agent_message_chunk",
892
+ content: { type: "text", text: "Reply" },
893
+ },
894
+ ])
895
+ }),
896
+ ),
897
+ ),
898
+ )
899
+
624
900
  it.effect("rejects unknown models and methods", () =>
625
901
  withStore(
626
902
  Effect.scoped(
package/src/Acp.ts CHANGED
@@ -29,7 +29,8 @@ import type * as Model from "effect/ai/Model"
29
29
  import * as Prompt from "effect/ai/Prompt"
30
30
  import * as HttpClient from "effect/http/HttpClient"
31
31
  import * as KeyValueStore from "effect/persistence/KeyValueStore"
32
- import type * as Agent from "./Agent.ts"
32
+ import * as Agent from "./Agent.ts"
33
+ import * as AgentExecutor from "./AgentExecutor.ts"
33
34
  import type * as AgentOutput from "./AgentOutput.ts"
34
35
  import * as Compaction from "./Compaction.ts"
35
36
  import * as Image from "./Image.ts"
@@ -125,6 +126,10 @@ export class SessionRecord extends Schema.Class<SessionRecord>(
125
126
  model: Schema.String,
126
127
  thoughtLevel: ThoughtLevel,
127
128
  history: Prompt.Prompt,
129
+ /**
130
+ * The agent's executor environment, such as its working directory.
131
+ */
132
+ environment: Schema.optionalKey(AgentExecutor.ExecutionEnvironment),
128
133
  }) {}
129
134
 
130
135
  /**
@@ -273,6 +278,7 @@ const historyUpdates = (history: Prompt.Prompt) => {
273
278
  for (const message of history.content) {
274
279
  if (message.role !== "user" && message.role !== "assistant") continue
275
280
  if (Compaction.isSummaryMessage(message)) continue
281
+ if (Agent.isInjectedMessage(message)) continue
276
282
  const sessionUpdate =
277
283
  message.role === "user" ? "user_message_chunk" : "agent_message_chunk"
278
284
  for (const part of message.content) {
@@ -297,6 +303,7 @@ interface Session {
297
303
  readonly agent: Agent.Agent
298
304
  running: Fiber.Fiber<void, unknown> | undefined
299
305
  toolCalls: number
306
+ readonly announcedSubagents: Set<number>
300
307
  }
301
308
 
302
309
  /**
@@ -311,6 +318,10 @@ export const make = Effect.fnUntraced(function* <RAgent, RModel>(
311
318
  KeyValueStore.KeyValueStore | Scope.Scope
312
319
  > {
313
320
  const kvs = yield* KeyValueStore.KeyValueStore
321
+ const { contextWindow } = yield* Compaction.CompactionConfig
322
+ const size = Number.isFinite(contextWindow)
323
+ ? Math.max(0, Math.round(contextWindow))
324
+ : 0
314
325
  // Optional: `https:` images need an HttpClient and `file:` images need a
315
326
  // FileSystem and Path. Inline (base64) images work without either. When a
316
327
  // service is missing the corresponding block is rejected, never fetched.
@@ -372,6 +383,7 @@ export const make = Effect.fnUntraced(function* <RAgent, RModel>(
372
383
  model: session.model,
373
384
  thoughtLevel: session.thoughtLevel,
374
385
  history: session.agent.history.current,
386
+ environment: session.agent.environment.current,
375
387
  }),
376
388
  )
377
389
  .pipe(
@@ -386,6 +398,9 @@ export const make = Effect.fnUntraced(function* <RAgent, RModel>(
386
398
  ) {
387
399
  const agent = yield* Scope.provide(options.makeAgent(record.cwd), scope)
388
400
  MutableRef.set(agent.history, record.history)
401
+ if (record.environment !== undefined) {
402
+ MutableRef.set(agent.environment, record.environment)
403
+ }
389
404
  const session: Session = {
390
405
  id,
391
406
  cwd: record.cwd,
@@ -394,6 +409,7 @@ export const make = Effect.fnUntraced(function* <RAgent, RModel>(
394
409
  agent,
395
410
  running: undefined,
396
411
  toolCalls: 0,
412
+ announcedSubagents: new Set(),
397
413
  }
398
414
  sessions.set(id, session)
399
415
  return session
@@ -652,6 +668,10 @@ export const make = Effect.fnUntraced(function* <RAgent, RModel>(
652
668
  case "Usage":
653
669
  return update(session.id, {
654
670
  sessionUpdate: "usage_update",
671
+ used: Number.isFinite(part.contextTokens)
672
+ ? Math.max(0, Math.round(part.contextTokens))
673
+ : 0,
674
+ size,
655
675
  usage: {
656
676
  inputTokens: Math.max(
657
677
  0,
@@ -669,20 +689,25 @@ export const make = Effect.fnUntraced(function* <RAgent, RModel>(
669
689
  `Retrying after error: ${part.error.message}\n`,
670
690
  ),
671
691
  })
672
- case "SubagentStart":
692
+ // A background agent's tool call stays in progress until the agent
693
+ // finishes, while the parent keeps taking turns.
694
+ case "SubagentStart": {
695
+ const announced = session.announcedSubagents.has(part.id)
696
+ session.announcedSubagents.add(part.id)
673
697
  return update(session.id, {
674
- sessionUpdate: "tool_call",
698
+ sessionUpdate: announced ? "tool_call_update" : "tool_call",
675
699
  toolCallId: `subagent-${part.id}`,
676
700
  title: `delegate: ${part.prompt.replace(/\s+/g, " ").trim().slice(0, 120)}`,
677
701
  kind: "think",
678
702
  status: "in_progress",
679
703
  rawInput: { prompt: part.prompt },
680
704
  })
705
+ }
681
706
  case "SubagentComplete":
682
707
  return update(session.id, {
683
708
  sessionUpdate: "tool_call_update",
684
709
  toolCallId: `subagent-${part.id}`,
685
- status: "completed",
710
+ status: part.status === "completed" ? "completed" : "failed",
686
711
  rawOutput: part.summary,
687
712
  content: [{ type: "content", content: textContent(part.summary) }],
688
713
  })
@@ -762,6 +787,11 @@ export const make = Effect.fnUntraced(function* <RAgent, RModel>(
762
787
  model: modelId,
763
788
  thoughtLevel: record.value.thoughtLevel,
764
789
  history: record.value.history,
790
+ // The stored environment belongs to the stored cwd.
791
+ ...(record.value.environment !== undefined &&
792
+ (cwd === undefined || cwd === record.value.cwd)
793
+ ? { environment: record.value.environment }
794
+ : {}),
765
795
  }),
766
796
  )
767
797
  if (replay) {
@@ -40,8 +40,14 @@ type Message = {
40
40
 
41
41
  const capabilities = new AgentExecutor.Capabilities({
42
42
  toolsDts: "",
43
- agentsMd: Option.none(),
43
+ environment: new AgentExecutor.ExecutionEnvironment({
44
+ cwd: "/",
45
+ baseline: [],
46
+ seen: [],
47
+ }),
48
+ instructions: Option.none(),
44
49
  supportsSearch: false,
50
+ tools: [],
45
51
  skills: [],
46
52
  })
47
53
 
package/src/Agent.test.ts CHANGED
@@ -18,13 +18,20 @@ import type * as Response from "effect/ai/Response"
18
18
  import * as ResponseIdTracker from "effect/ai/ResponseIdTracker"
19
19
  import * as Agent from "./Agent.ts"
20
20
  import * as AgentExecutor from "./AgentExecutor.ts"
21
+ import { subagentsUnavailable } from "./AgentTools.ts"
21
22
  import * as AgentOutput from "./AgentOutput.ts"
22
23
  import * as Compaction from "./Compaction.ts"
23
24
 
24
25
  const capabilities = new AgentExecutor.Capabilities({
25
26
  toolsDts: "",
26
- agentsMd: Option.none(),
27
+ environment: new AgentExecutor.ExecutionEnvironment({
28
+ cwd: "/",
29
+ baseline: [],
30
+ seen: [],
31
+ }),
32
+ instructions: Option.none(),
27
33
  supportsSearch: false,
34
+ tools: [],
28
35
  skills: [],
29
36
  })
30
37
 
@@ -114,7 +121,10 @@ describe("Agent", () => {
114
121
  conversationMode: true,
115
122
  executor: makeExecutor(({ onTaskComplete }) =>
116
123
  Stream.fromEffect(
117
- onTaskComplete("done from taskComplete").pipe(Effect.as("")),
124
+ onTaskComplete("done from taskComplete").pipe(
125
+ Effect.as(""),
126
+ Effect.orDie,
127
+ ),
118
128
  ),
119
129
  ),
120
130
  streamText: () =>
@@ -162,6 +172,7 @@ const runAgentCollect = (options: {
162
172
  readonly compaction?: Partial<Compaction.CompactionConfigService> | undefined
163
173
  readonly prompt?: string | undefined
164
174
  readonly turnTimeout?: Duration.Duration | undefined
175
+ readonly maxRetries?: number | undefined
165
176
  readonly respond: ScriptedResponse
166
177
  }) =>
167
178
  Effect.scoped(
@@ -243,6 +254,9 @@ const runAgentCollect = (options: {
243
254
  Agent.TurnTimeout,
244
255
  options.turnTimeout ?? Duration.minutes(5),
245
256
  ),
257
+ options.maxRetries === undefined
258
+ ? Layer.empty
259
+ : Agent.MaxRetries.layer(options.maxRetries),
246
260
  ),
247
261
  ),
248
262
  Effect.exit,
@@ -467,14 +481,17 @@ it.layer(localExecutorLayer, { excludeTestServices: true })(
467
481
  AgentExecutor.AgentExecutor.use((executor) =>
468
482
  Effect.gen(function* () {
469
483
  const summaries: Array<string> = []
484
+ const { environment } = yield* executor.capabilities
470
485
  const output = yield* executor
471
486
  .execute({
487
+ environment,
488
+ onEnvironment: () => Effect.void,
472
489
  script: `try { await ${invocation}; console.log("ACCEPTED") } catch (error) { console.log("REJECTED:", String(error)) } console.log("SCRIPT_SURVIVED")`,
473
490
  onTaskComplete: (summary) =>
474
491
  Effect.sync(() => {
475
492
  summaries.push(summary)
476
493
  }),
477
- onSubagent: () => Effect.succeed(""),
494
+ subagents: subagentsUnavailable,
478
495
  onImage: () => Effect.void,
479
496
  })
480
497
  .pipe(Stream.mkString)
@@ -1164,7 +1181,7 @@ describe("Agent overflow compaction", () => {
1164
1181
  })
1165
1182
  assert.isTrue(Exit.isFailure(exit), "turn must fail")
1166
1183
  const error = Option.getOrThrow(Exit.findErrorOption(exit))
1167
- assert.strictEqual(error._tag, "AiError")
1184
+ assert(error._tag === "AiError")
1168
1185
  assert.strictEqual(error.reason._tag, "InvalidRequestError")
1169
1186
  assert.strictEqual(calls.length, 3)
1170
1187
  }),
@@ -1285,3 +1302,91 @@ describe("Agent overflow compaction", () => {
1285
1302
  }),
1286
1303
  )
1287
1304
  })
1305
+
1306
+ describe("Agent retry limit", () => {
1307
+ it.live(
1308
+ "fails with RetriesExhausted once model errors and turn timeouts reach MaxRetries",
1309
+ () =>
1310
+ Effect.gen(function* () {
1311
+ const { exit, calls } = yield* runAgentCollect({
1312
+ maxRetries: 2,
1313
+ turnTimeout: Duration.millis(100),
1314
+ respond: (_call, index) => {
1315
+ switch (index) {
1316
+ case 0:
1317
+ return Stream.fail(providerError)
1318
+ case 1:
1319
+ return Stream.fromEffect(Effect.never)
1320
+ case 2:
1321
+ return Stream.fail(providerError)
1322
+ default:
1323
+ return Stream.die(
1324
+ `unexpected model call #${index}: must stop retrying`,
1325
+ )
1326
+ }
1327
+ },
1328
+ }).pipe(Effect.timeout("3 seconds"))
1329
+ assert.strictEqual(calls.length, 3)
1330
+ const error = Option.getOrThrow(Exit.findErrorOption(exit))
1331
+ assert(error._tag === "RetriesExhausted")
1332
+ assert.strictEqual(error.retries, 2)
1333
+ assert.strictEqual(error.error.reason._tag, "InternalProviderError")
1334
+ }),
1335
+ )
1336
+ })
1337
+
1338
+ describe("Agent history events", () => {
1339
+ it.effect(
1340
+ "rebuild the agent's history from the history it was resumed with",
1341
+ () =>
1342
+ Effect.gen(function* () {
1343
+ const { exit, outputs, history } = yield* runAgentCollect({
1344
+ history: fatHistory,
1345
+ compaction: {
1346
+ contextWindow: 12_000,
1347
+ reserveTokens: 2_000,
1348
+ keepRecentTokens: 9_000,
1349
+ },
1350
+ executor: makeExecutor(() => Stream.succeed("listed")),
1351
+ respond: (call, index) => {
1352
+ switch (index) {
1353
+ case 0:
1354
+ assertSummarizerCall(call)
1355
+ return Stream.fromIterable(text("SUMMARY"))
1356
+ case 1:
1357
+ return Stream.fromIterable([
1358
+ toolCall("call-3", "console.log(await ls('.'))"),
1359
+ finish(500),
1360
+ ])
1361
+ case 2:
1362
+ return Stream.fromIterable([
1363
+ ...text("done"),
1364
+ finish(600, "stop"),
1365
+ ])
1366
+ default:
1367
+ return Stream.die(`unexpected model call #${index}`)
1368
+ }
1369
+ },
1370
+ })
1371
+ assert.deepStrictEqual(exit, Exit.succeed("done"))
1372
+
1373
+ let replayed: Array<Prompt.Message> = [...fatHistory.content]
1374
+ for (const output of outputs) {
1375
+ if (output._tag === "HistoryAppended") {
1376
+ replayed.push(...output.messages)
1377
+ } else if (output._tag === "HistoryReplaced") {
1378
+ replayed = [...output.messages]
1379
+ }
1380
+ }
1381
+ assert.strictEqual(outputsOfTag(outputs, "HistoryReplaced").length, 1)
1382
+ assert.strictEqual(
1383
+ promptJson(Prompt.fromMessages(replayed)),
1384
+ promptJson(
1385
+ Prompt.fromMessages(
1386
+ history.content.filter((message) => message.role !== "system"),
1387
+ ),
1388
+ ),
1389
+ )
1390
+ }),
1391
+ )
1392
+ })